Unsupervised Change Detection of Multi-temporal Satellite Images using PSO-based Image Thresholding Technique

Timely and accurate detection of changes in land use/ cover is important for land planning and management. Remote sensing images have been primary sources for change detection in recent decades. Due to its simplicity, thresholding of difference image is a popular method for change detection. The traditional thresholding methods such as Otsu are based on exhaustive search, so that they are time consuming. Since these methods are mainly developed for one-dimensional problems, the computation time grows exponentially with the number of thresholds when these methods are extended to be used for multi-dimensional problems. If thresholding is supposed to be as an optimization problem, optimization methods can potentially decrease the computation time. In this paper, a fast, simple and effective multi-dimensional image thresholding technique based on Particle Swarm Optimization (PSO) method is presented. This technique calculates the optimal threshold values by maximizing the Otsu objective function and minimizing the inter-class variance objective function. The proposed method has been implemented on two multispectral and multi-temporal datasets. The first dataset includes a couple of images acquired by the TM sensor taken form south islands of Aurmia Lake (Iran) in Jun 1984 and July 2010, respectively. The second dataset is obtained from a couple of images acquired by the same sensor on the Khodafarin dam (Iran) in July 2000 and July 2009, respectively. In order to evaluate the proposed method, the computational time and change detection accuracy were computed. In addition, statistical test was carried out in order to evaluate the robustness of the developed method. The experimental results show that the proposed PSO-based multi-dimensional thresholding method could provide optimum thresholds values by decreasing 98% and 15% of the time complexity compared with the most widely used Otsu and inter-class variance-based thresholding methods.